optimizing upf performance using smartnic offload · 2020. 11. 24. · 1 mavenir.com solution brief...

Copyright © Mavenir 2020. All rights reserved.

1 mavenir.com

S O L U T I O N B R I E F

OPTIMIZING UPF PERFORMANCE USING SMARTNIC OFFLOAD

5G is bringing about a transformation in connectivity. 5G not only enhances mobile broadband experiences but also supports many different verticals like AR/VR, Automotive, Energy, and Public Safety. The 5G landscape will see enablement of services across verticals with varying performance needs. This sets very challenging and demanding requirements on latency and throughput for a better experience. Since the User Plane Function (UPF) is in the path of processing all network data, it plays a key role in the 5G network and the ability to realize the ambitious vision of supporting low latency and high throughput requirements.

Mavenir’s User Plane Function

Mavenir’s User Plane Function (UPF) handles the critical operation of processing subscriber traffic between the RAN (Radio Access Network) and the DN (Data Network). Mavenir’s design is flexible for various deployment options that are based on operator-specific needs and use-cases. This can be either as co-located with the other Core Network functions or deployed at the edge of the network aggregation point.

Mavenir’s UPF is a cloud-native, highly optimized packet processing design with configurable hardware offloading capabilities, and automation for edge deployments.

Figure 1: Mavenir’s 5G UPF

KEY FEATURES OF USER PLANE FUNCTION

• Container based Network Function

• VPP based packet processing architecture

• Optimized software for packet processing

• 3GPP Release 15 and Release 16 support

• Combo Node (UPF + PGW-U + SGW-U + GGSN-U)

• EPS Interworking • Support IPv4, IPv4v6, and

IPv6 PDUs • Uplink Classifier and

Branching Point support • NRF based registration and

de-registration • N4/PFCP Session Management • Lawful Intercept Support • Transport level marking (DSCP) • Supports N3, N6, N9 interfaces • Supports IPv4, IPv4v6,

IPv6 PDUs • IPv6 Multi-Homing • N6 Gi LAN services such as

CGNAT and Firewall Support • Application Awareness and

Deep Packet Inspection support up to layer 7

• Session Flow Data Records and Transaction logs

• Tethering Detection and Rate Limiting support

• Hardware offloading support • Roaming and IPUPS Support


2 mavenir.com


While meeting very high-performance demands and low hardware footprint, Mavenir UPF also delivers on deployment automation and ease of operational maintenance through centralized management system, reducing total cost of ownership.

To reduce CAPEX for 5G, Mavenir UPF provides 5G UPF and PGW-U/SGW-U/GGSN-U support, delivering a unique combination of high performance and capability to support 2G/3G, 4G and 5G subscribers in a single network function.

Achieve 524 Gbps throughput with 50% reduction in compute costs This is achieved with UPF Packet Processing offload using SmartNICs (Smart Network Interface Controller) Before SmartNIC offload: No packet processing activities or features like GTPU encapsulation/ decapsulation and flow tagging are offloaded.

Packet Processing Cores : 32, Avg. CPU Utilization : 78%, Throughput : 524 Gbps

Figure 2: Mavenir’s 5G UPF

Figure 3: Before SmartNIC offload


3 mavenir.com


Mavenir UPF SmartNIC offload capability offers efficient distribution of the user plane packets across multiple CPUs. Other features offloaded to SmartNIC are GTPU encapsulation and decapsulation, and flow tagging.

After SmartNIC offload: Packet processing activities or features like GTPU encapsulation/ decapsulation and flow tagging are offloaded to SmartNIC.

The offloading capability delivers the following benefits:

• Significantly reduces latency/jitter, avoids inter CPU handover leading to a Higher Throughput • Since there is no cache miss, it consequently Reduces Overheads, Improves Latency and

Performance

Mavenir UPF’s unique design allows efficient resource management of available CPU that further enables:

• Ability to handle varying types of user traffic with different packet sizes • Ease of handling higher peak data rates with NSA option 3x and 5G SA peak data rates • Higher system throughput capacity without impacting the number of subscriber sessions being

handled by the UPF

Packet Processing Cores: 16 (reduced by 50%), Avg. CPU Utilization: 72% (drops by 6%), Throughput: 524 Gbps

Figure 4: After SmartNIC offload


4 mavenir.com


The UPF Packet Processing is based on FD.IO’s Vector Packet Processing (VPP) technology that provides the benefits of a programmable data plane:

• Enables diverse ways of packet forwarding and supports various formats and protocols making its management easy and flexible

• Allows a programmable packet processing pipeline that further enables deployment of the porting solution on hardware platforms such as x86, standard NIC, SmartNIC

The platform supports various deployments such as edge, far edge, central, which can be scaled to support required performance needs

Software optimization done by Mavenir using VPP implementation includes benefits such as:

• Optimized use of CPU’s instruction cache, memory footprint reduction to fit L2/L3 cache for best performance

• Use of software pre fetch to hide memory access latency

UPF Benchmarking

Mavenir conducted a benchmarking test to quantify the UPF performance improvements by offloading the following UPF functions to the SmartNIC: GTP RSS, GTP Encapsulation, GTP Decapsulation and Flow Tagging. We conducted a test with 3 configurations.

Test Configuration 1: There is no offload in this scenario.

Test Configuration 2: Only RSS is offloaded, Flow Tagging and GTP Encapsulation/Decapsulation are not.

Test Configuration 3: RSS, Flow Tagging and GTP Encapsulation/Decapsulation are offloaded

Test Configuration 1 Test Configuration 2 Test Configuration 3

RSS Offload No Yes Yes

Flow Tagging No No Yes

GTP Encapsulation/Decapsulation No No Yes

Packet Processing Cores 32 22 16

Total Throughput (Gbps) 524 524 524

Average CPU Utilization (%) 78% 76% 72%

Downlink PPS (in millions) 57.5 57.5 57.5

Uplink PPS (in millions) 28.9 28.9 28.9

Average Packet Size (bytes) 743 743 743


5 mavenir.com


Test Lab Configuration 1. Dell R740 server with two Intel® Xeon® Gold 6248R processors, 384 GB RAM and a 1.6 TB SSD

Drive. § R740 Dell Server

Intel® Xeon® Gold 6248R 3.0G, 24C/48T, 10.4GT/s, 35.75M Cache, Turbo, HT (205W) DDR4-2933 5-PCIe3x16, 384GB RAM, TB 7.2K RPM NLSAS 12Gbps 512n 3.5in Hot-plug Hard Drive RAID

4x900 GB SSD (total 3.6TB) 2. The server featured Dual-Port 100G NVIDIA Mellanox ConnectX-6 NICs.

§ A total of 6 NICs of these were installed on the server 3. Software included Containerd v1.2.10, Kubernetes v1.17.3, DPDK 19.11, and VPP 19.01 4. The testing environment was set up with 50,000 emulated subscribers 5. 2PDR, 2URR and 2QER per session were enabled; L7 DPI were not enabled 6. Total of 6 UPF PODs where deployed on the same Dell server. Per the picture below UPF1, UPF2,

UPF3 and UPF6 were assigned 16 PCIe slots each. UPF4 and UPF6 were assigned 8 PCIe slots each.

KEY FINDING:

50% reduced compute cost for the same throughput of 524 Gbps.

Mavenir was able to demonstrate 524 Gbps of throughput on a single server using 16 packet processing cores. As per the figure below, Mavenir was able to reduce the UPF footprint from 32 FastPath (FP) pCores when no offload was used to 16 FastPath (FP) pCores when GTP RSS, GTP Encapsulation, GTP Decapsulation and Flow Tagging was offloaded to the SmartNIC. That is a 50% of FastPath (FP) pCore savings in order to achieve the same 524 Gbps throughput

Figure 5: UPF performance results


6 mavenir.com


Mavenir’s UPF Architecture

Mavenir’s solution is modular, extensible, scalable, service-based, and highly optimized to meet the higher performance requirements for 5G systems.

UPF Architecture Layers

• Management Layer: Performs application configuration management and OAM procedures • Session Layer: Performs UPF session handling and interfaces with other control plane

elements • UPF Agent: Provides fast path services to slow path applications • Policy Layer: Provides fast path application logic of policy lookup, execution, and flow

offloading • Flow Layer: Terminates N3, N6/N9, S1/S5/S8/Gn/Gp and SGi Interfaces. Runs Graph nodes

for IPv4/v6, GTPU, UPF Session Lookup, PDR rule match & Action handling, QoS Enforcement, DPI, CGNAT

KEY ARCHITECTURE COMPONENTS

The Vector Packet Processing framework implements plug and play graph nodes thus enabling programmability and faster time to market

The framework is optimized to process high data rates and support a large number of flows with varying packet sizes

It supports receiver side scaling to load balance traffic based on a various set of criterial elements such as source and destination address based load balancing

The UPF framework allows the offloading of packet processing components to SmartNICs

Figure 6: Mavenir's UPF Architecture


7 mavenir.com


Mavenir UPF Packet Processing Evolution

Mavenir’s UPF is evolving over multiple phases to deliver a high performance and cost optimized UPF. Over these phases, Mavenir is developing and enhancing its UPF leveraging the benefits of Software acceleration techniques like DPDK and VPP, as well as Hardware optimization techniques like SmartNIC and Switch Fabric Offload.

UPF Differentiators Flexible Architecture Value-add Services § Flexible hardware offload § 2G/3G/4G/5G combo node

support § UPF internetworking with any

vendor’s SMF via N4 § Integrated UPF and CU-UP

§ VPP plug-ins for GLAN services (i.e. DPI, Tethering, CGNAT, FW)

§ Policy Rules and Enforcement

Performance Redundancy & Scalability § Containerized UPF running on

bare metal § 524Gps on 16 packet

processing pCores § No east-west traffic, single

hop

§ M : N redundancy § Fast Switchover during HA § K8s and Application KPI

based Scaling

Figure 7: Mavenir's UPF evolution

Figure 8: UPF Differentiators


8 mavenir.com


CONCLUSION

Achieve Lower TCO using Mavenir’s UPF

TCO is a crucial element in the transition to 5G. While multiple aspects influence the TCO towards 5G, there are specific factors of Mavenir UPF implementation that optimize the cost as noted below:

a. Mavenir UPF’s cloud-native design renders a smaller hardware footprint. The containerization of the UPF software offers increased flexibility, efficient resource utilization and elimination of dependency on hardware.

b. Cloud-native applications are managed dynamically on a proven Kubernetes based container-orchestration platforms that offer hardware decoupling that is essential for deployment automation.

c. Less hardware and efficient network management systems improve operational resiliency and reduce complexity, thus lowering costs.

An Open, Highly Flexible Solution using Cloud-native Principles

Mavenir is a leader in accelerating and redefining network transformation for Service Providers, by offering a comprehensive product portfolio across every layer of the network infrastructure stack. From 4G and 5G application/service layers to 5G core and EPC and open RAN – Mavenir leads the way in evolved, cloud-native networking solutions enabling innovative and secure experiences for end users.

Mavenir ’s solutions are deployed at 250+ CSPs who cover about 4B subscribers. Mavenir ’s platform enables these CSPs to successfully deliver next generation v ision today and realize new revenue streams and operational eff iciencies

Mavenir ’s containerized packet core solution leverages proven industry tools, software, and best practices to provide an open and highly flexible solution. Using general purpose servers and standard operating systems (Linux), Mavenir packet core solution including UPF is designed for maximum interoperabil ity, allowing seamless integration with third-party tools for extended functionality. Fully automated Life Cycle Management (LCM) and scalability based on Kubernetes integration. Continuous Integration and Continuous Delivery (CI/CD) software pipeline. The system can scale down to meet far edge deployment use case requirement. The UPF should be dynamically deployable and granularly scalable. This can only be achieved through cloud-native web scale deployment methodologies, UPF as a container can be remotely instantiated and maintained further reducing the edge deployment time and operation costs.

Maximizing Investments

To help optimize technology investments, Mavenir and its partners offer complete solutions that inc lude professional serv ices, technical support, and education.

For more information, contact a Mavenir sales partner or visit www.mavenir.com Rev01-20-11-20


9

mavenir.com


LIST OF ACRONYMS

Acronym Explanation

3GPP 3rd Generation Partnership Project

5G 5th Generation

AR/VR Augmented Reality/Virtual Reality

CGNAT Carrier Grade NAT (Network Address Translation)

DN Data Network

DPDK Data Plane Development Kit

DPI Deep Packet Inspection

EPS Evolved Packet System

GGSN Gateway GPRS Support Node

GTP GPRS Tunneling Protocol

LAN Local Area Network

NRF Network Repository Function

NSA Non-Standalone

OAM Operation and Maintenance

PDU Packet Data Unit

PFCP Packet Forwarding Control Protocol

PGW Packet Gateway

PPS Packets Per Second

QoS Quality of Service

RAN Radio Access Network

RSS Rich Site Summary

SGW Serving Gateway

UPF User Plane Function

VPP Vector Packet Processing

optimizing upf performance using smartnic offload · 2020. 11. 24. · 1 mavenir.com solution brief...

Documents